Research On Vehicle Re-identification Based On Deep Learnin

Vehicle re-identification technology is also known as cross-camera tracking technology for vehicles.The aim of vehicle re-identification is to find all the vehicles identical to the vehicle to be queried from the surveillance video images captured by non-overlapping cameras.In recent years,the vehicle Re-ID technology has attracted more and more attention from academics and industries,and it has a wide range of applications in intelligent transportation systems.Vehicles not only have a wide variety,but also different forms,with typical "inter-class small differences,intra-class differences " challenge,such as different vehicles due to the same model,color and even manufacturer,their appearance differences are small,while the same vehicle due to the different traffic scenes,their appearance differences are huge.In addition,in combination with the drastic change of shooting viewpoint when the vehicle is moving across the camera,the challenge of vehicle re-identification is intensified.To address the challenges faced above,this paper studies the vehicle re-identification task based on deep learning ideas,the main work is as follows:(1)For the challenge that many vehicles have near-parallel appearance in complex traffic scenes leading to the decrease of vehicle re-identification accuracy,this paper proposes a three-branch embedding network with part-aware ability and feature complementary learning.Among them,the global branch is used to extract the global features of vehicles.The local branch goes to learn subtle but discriminative vehicle part features in the image through an attention mechanism.In order to accommodate the constant dynamic changes in vehicle appearance during cross-camera motion,the complementary branches are designed to extract more complete structural features and multi-granularity features.Finally,the features of the three branches are integrated into a unified framework for end-to-end learning,which not only obtains richer and more complete vehicle features,but also adapts to the ever-changing critical local information in vehicle images.(2)To address the low generalization ability of local feature learning,this paper proposes a view-aware part enhancement module that adaptively enhances salient local feature learning by means of weakly supervised attention learning.Compared with the fixed local information obtained by strongly supervised manual local annotation-based methods,the model in this paper not only makes it easier to locate more local regions within the viewpoint,but also reduces the cost of extensive and expensive annotation and enhances the local learning generalization ability.(3)In order to reduce the problem of huge viewpoint bias in vehicle cross-view re-identification,this paper proposes a common-view attentive promote module,which calculates attention weights based on the similarity between viewpoints of matched vehicles,so that feature learning under the same view is enhanced and feature learning from different viewpoints is suppressed,thus enhancing the difference in viewpoint features and robustness to drastic viewpoint changes...